- I've made a 7B model run on a laptop GPU that had no business running one.SNKT
- I've rebuilt RAG from open-source parts — embeddings, chunking, rerankers, quantization.FWUMP
- I've put guardrails on a medical AI so it never plays doctor.NOPE
- I've reconciled five marketplaces before breakfast.SYNC
- Zero data-loss incidents. Not once.KNOCK
ALRIGHT. LET'S DO THIS ONE LAST TIME.
KIRTAN
TANK
AI ENGINEER · BACKEND · SURAT, INDIA
I learned this from the bottom up — open-source models, on my own hardware.
Two years building AI that has to work on a Tuesday, not just in a demo.
YOU KNOW
THE REST
2020 – 2024 · GTU (RNGPIT)
EDUCATION
B.E. Computer Science & Engineering
2020 – 2024
GUJARAT TECHNOLOGICAL UNIVERSITY
(RNGPIT)
Minor specialisation in Artificial Intelligence and Machine Learning. The degree got me in the room. Everything past this page is what I did once I was there.
LLM Engineering
HF Agentic AI
HF MCP Course
NEKTAR P.L. · 2024
EARLY AI WORK
Where I learned RAG the hard way — from the bottom up, on a laptop.
Make the AI read the documents. And answer from them.
Every tutorial used the OpenAI API, which had just stopped being free — so I built it out of open-source parts instead. Embeddings, chunking, rerankers. Benchmarked FAISS, Milvus and Chroma for latency and recall. Then the model I needed didn't fit in 12 GB, and quantization became the whole problem.
LangChainLlamaIndex
FAISSBGE / GTE
GPTQ · AWQ · QAT → GPTQ
IT RAN.
Quantized Zephyr-7B, on a 12 GB laptop GPU
−65%inference memory
+30%answer relevance
1developer on it
WHAT I BUILD WITH
THE STACK
GENERATIVE AI
RAGAgentic workflows
Guardrail designEvals & benchmarking
Quantization
GPT-4oClaude
GeminiLlama 4
RAG & TOOLING
Hybrid searchChunking strategy
Rerankers
LangChainLlamaIndex
smolagentsFAISS
MilvusPinecone
FULL-STACK
PythonTypeScript
Next.jsNode.js
FastAPIPostgreSQL
PrismaSupabase
Frappe / ERPNext
DEVOPS
DockerNginx
VPS / LinuxGitHub Actions
VercelFrappe Cloud
DR backups
APR 2024 → PRESENT
MY WORK
Client names are under NDA. The problems aren't.
AGENTIC RAG
US real-estate platformHuman-in-the-loop agentic RAG over live property APIs and FAQ documents — and it knows when to stop and ask a human.
LangChainLlamaIndex
FAISSGPT-4o
+25%response accuracy
MEDICAL GUARDRAILS
Dermatology clinic groupHIPAA-compliant image analysis, built around what it must never do: give medical advice.
FastAPIGPT-4o
GeminiClaude
HIPAA
4models benchmarked, 1 shipped
INVOICE INTAKE
Global industrial manufacturerAn LLM reads purchase invoices and files them straight into SharePoint and Excel. Nobody types a number.
n8nClaude API
SharePoint
−80%manual data entry
RECONCILIATION
D2C brand · 5 marketplacesA Frappe back office pulling orders from five marketplaces — reconciled, then tracked all the way to revenue. Now: manufacturing.
FrappeERPNext
Marketplace APIs
15 hrs/week saved
WORKSHOP SAAS
Automotive service platformGST-compliant sequential invoicing, an edit-approval workflow and owner analytics. Mine from the first commit.
Next.js 14Prisma
PostgreSQLDocker
−60%invoice prep
~99.5%uptime
BLOCKCHAIN × AI
US fintech · code reviewCode review and polishing on a system where an immutable on-chain ledger held the audit trail — and AI agents triaged incoming transactions: auto-clear the small ones, park anything over threshold for human approval. Same shape as the agentic RAG above, with money on the line.
SoliditySmart contracts
MetaMaskEVM / gas
ALSO SHIPPED
Five more, in briefStaged IVF report summaries for a fertility clinic. A certification portal and LMS. A field-service platform that replaced paper. A multi-agent concierge. And an OCR pipeline that reads what people scribbled, then GPT-3.5 to clean it up.
Frappe LMSLlamaIndex Workflows
FlowiseAIPaddleOCR
GPT-3.5-turboSupabase
~88%char accuracy
NONE OF IT MATTERS IF THE BACKEND IS A HOUSE OF CARDS.
So I got good at the boring parts too — schemas, scheduled jobs, retries, RBAC, and the error handling that decides whether a bad API day becomes an incident.
Turns out the AI was never the hard part.
SIDE QUESTS
My other personal projects.
LIVE
LEDGER
An agent harness over an MCP tool server — refunds need code authorisation and a human approval, and a red-team measured what prompt injection does with and without them.
MCPHarness eng.
Context eng.
SEE IT RUN →
LIVE
GAIA AGENT
Multimodal agent for the GAIA benchmark — reasoning, web search, page visits and audio transcription.
OSS-120BWhisper
Llama 4 Maverick
TRY THE DEMO →
EVENT PLANNER AI
A CodeAgent on Hugging Face smolagents that plans events through web search and tool automation — with tracing, because agents lie.
smolagentsGroq
LangFuse
SOURCE →
I'M GOING DEEPER INTO
AI ENGINEERING — ON A BACKEND
THAT DOESN'T FLINCH.
Two years in, the thing I'm best at is the seam between them.
THE LONG VERSION How I learned this with no API budget → Origin story · ~7 minEND OF ISSUE
AI & backend engineer at RedSoft Solutions, working from Surat, India. If you've got a system that needs to be smart and stay up, I'd like to hear about it.